Semantic analysis method and system based on history teaching

Abstract

The invention discloses a semantic analysis method and system based on history teaching, and relates to the technical field of artificial intelligence. Comprising the steps of obtaining text data, and generating a standardized text through character-level cleaning and grammar rule matching; performing semantic extraction on the standardized text, identifying a historical event entity, a character relationship and a timeline grammar structure, and generating a structured semantic unit; constructing a dynamic knowledge graph based on the structured semantic unit, and defining event-character-time three-dimensional node attributes and a causal-time sequence-space edge relationship; converting the knowledge graph into multi-modal visual results such as a time axial flow type display result, a relational network topological graph and an event scene recovery graph; and collecting user interaction feedback data, optimizing knowledge graph node weights through a semantic similarity algorithm, and iteratively updating the semantic analysis model. The problems of complex text processing and weak knowledge association in history teaching are solved, and the accuracy and interactivity of semantic analysis are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and relates to language semantic understanding analysis and knowledge extraction. Specifically, it is a semantic analysis method and system based on historical teaching. Background Art

[0002] In the field of historical teaching, traditional text analysis methods mainly rely on manual annotation and keyword matching, and have the following technical defects: Serious fragmentation in historical text processing: Existing general text processing technologies (such as the RAG text segmentation method in the published patent 2: CN120353880A) focus on fixed-length or semantic boundary segmentation. However, historical texts contain a large number of ancient Chinese sentence patterns (such as the judgment sentence “...zhe...ye”), special time expressions (such as Heavenly Stems and Earthly Branches chronology, reign title chronology), and proper nouns (such as alternative names for people, evolution of place names). The fixed segmentation mode is prone to semantic breakage (such as splitting the causal relationship between the “Opium War - Westernization Movement” into independent text blocks), and cannot completely retain the relevance of historical events.

[0003] 2. Insufficient accuracy and single dimension in semantic extraction: Existing retrieval and consultation systems (such as the published patent 1: CN120086428A) achieve user demand profiling through multi-modal information extraction, but do not design a dedicated semantic model for the particularity of historical texts: Event recognition only relies on keyword matching, making it difficult to distinguish between the “First Sino-Japanese War” (historical event) and the “Jiawu Year” (time expression); Person relationship extraction is limited to explicit associations (such as “leadership - being led”), and cannot capture implicit historical connections (such as the master-apprentice relationship between “Zeng Guofan - Li Hongzhang”); Timeline parsing lacks historical staging logic (such as the definition of fuzzy time intervals like “late Qing Dynasty” and “Republic of China”).

[0004] 3. Disconnection between knowledge display and teaching needs: The visualization methods of existing technologies (such as the sentiment analysis adjustment display in the published patent 1) focus on information presentation efficiency and do not consider the scenario-based needs of historical teaching; Lack of integrated display of the time axis and spatial relationship (such as the spatial mapping between the “Opium War” and the location where the “Treaty of Nanjing” was signed); Unable to dynamically optimize the knowledge structure through interactive feedback, resulting in fragmented understanding of the causal chain and person network of historical events by students.

[0005] 4. Poor adaptability of cross-domain knowledge transfer: The cross-domain retrieval technology of the published patent 1 and the dynamic chunking method of the published patent 2 are both oriented to general fields and do not design a knowledge organization framework for the four-dimensional association characteristics of “event - person - time - space” in the historical discipline, resulting in low structuralization of historical knowledge and unable to support in-depth teaching analysis (such as the multi-dimensional association analysis of the “social background and influence of the 1911 Revolution”).

[0006] To address the aforementioned issues, a semantic analysis technology that integrates historical grammar rules, dynamic knowledge graphs, and teaching feedback mechanisms is needed to provide intelligent tools to support history teaching. Summary of the Invention

[0007] This invention provides a semantic analysis method and system based on history teaching. Through standardized text processing, multi-dimensional semantic extraction, visualized knowledge display, and iterative optimization, it realizes the structured, associative, and contextualized analysis of history teaching texts, solving the shortcomings of traditional methods such as fragmented historical knowledge, low semantic extraction accuracy, and poor teaching adaptability.

[0008] To achieve the above objectives, the present invention provides: A semantic analysis method based on history teaching includes the following steps: S1: Obtain historical teaching text data, perform character-level preprocessing and grammar rule verification on the text data, and generate standardized historical text data; S2: Perform semantic extraction on standardized historical text data, identify historical event entities, relationships between people and timeline grammatical structures, and generate structured historical semantic units; S3: Construct a dynamic knowledge graph based on structured historical semantic units, defining three-dimensional node attributes of event-person-time and causal-temporal-spatial edge relationships; S4: Transform dynamic knowledge graphs into multimodal visualizations, presenting semantic analysis results through timeline streaming, relational network topology diagrams, and event scene reconstruction diagrams; S5: Collect user interaction feedback data, optimize the weight of knowledge graph nodes through semantic similarity algorithms, and iteratively update the semantic analysis model.

[0009] Further, step S1 includes the following steps: S11: Perform character-level cleaning on historical teaching text data to remove non-text symbols and repeated character sequences; S12: Perform grammatical verification on the cleaned text based on the historical grammar rule base, and mark the subject-verb-object structure and time adverbial clauses; S13: Convert the grammar-checked text into a standardized format, unifying the time representation and spelling of historical proper nouns.

[0010] Further, step S2 includes the following steps: S21: Use a bidirectional LSTM model to perform event entity recognition on standardized historical text data and extract event name, time of occurrence and location information; S22: Extract character entities and relation triples through the BERT pre-trained model to generate "character-action-object" semantic units; S23: Based on the dependency syntactic analysis of the timeline grammar structure, establish the mapping relationship of the event occurrence order and time interval.

[0011] Furthermore, step S3 includes the following steps: S31: Define the node attributes of the dynamic knowledge graph. The event nodes include type labels, importance scores, and literature sources; S32: Construct an edge relationship weight calculation model. The causal association weight is based on the event co-occurrence frequency, and the temporal dependence weight is based on the time interval decay function; S33: Optimize the structure of the knowledge graph through a graph neural network to eliminate isolated nodes and redundant edge relationships.

[0012] Furthermore, step S4 includes the following steps: S41: The time-axis flow display module arranges event nodes in timestamp order, supporting zooming and interval selection interactions; S42: The relationship network topology diagram uses a force-directed layout algorithm. The node size maps the event importance, and the edge thickness maps the relationship weight; S43: The event scene restoration diagram generates semi-realistic images in combination with entity attributes, and annotates key characters and event elements.

[0013] Furthermore, step S5 includes the following steps: S51: Collect user click heat, stay duration, and path trajectory data to generate an interaction feedback matrix; S52: Calculate the semantic similarity between the feedback matrix and the knowledge graph nodes, and adjust the node weights and display priorities; S53: Iteratively update the semantic extraction model through a reinforcement learning algorithm to optimize the historical event recognition accuracy.

[0014] Furthermore, the historical grammar rule library in step S12 includes: S121: Ancient Chinese special sentence pattern rules, including the "……者……也" structure of judgment sentences and the "为……所……" structure of passive sentences; S122: Historical document time expression rules, covering the automatic conversion of year number chronology, heavenly stem and earthly branch chronology, and Gregorian calendar; S123: Historical proper noun spelling rules, including variant names of people, evolution of place names, and norms for official titles.

[0015] Furthermore, the relationship triple extraction in step S22 includes: S221: Person-event association, marking the role and degree of influence of a person in an event; S222: Person-person association, distinguishing kinship, superior-subordinate relationship, and cooperation / opposition relationship; S223: Event-Location Association, records the precise geographical location and spatial extent of the event.

[0016] Furthermore, the generation of the event scene reconstruction map in step S43 includes: S431: Call the historical scene material library based on event node attributes to match era characteristics and regional styles; S432: Semi-realistic images are synthesized using a GAN network, with figures' clothing and objects conforming to historical research. S433: Overlay event element annotations, including timestamps, names of key figures, and summaries of core events.

[0017] A semantic analysis system for history teaching, comprising: The data processing module is used to acquire historical teaching text data, perform character-level cleaning and grammar rule matching on the text data, and generate standardized historical text data. The semantic analysis module performs multi-dimensional semantic extraction on standardized historical text data, including historical event entity recognition, person relationship extraction, and timeline grammar parsing, to generate structured historical semantic units. The knowledge graph construction module builds a dynamic knowledge graph based on structured historical semantic units. Graph nodes contain event types, person attributes, and timestamp information, and edge relationships include causal associations, temporal dependencies, and spatial mappings. The image display module converts dynamic knowledge graphs into multimodal visualization results, including timeline streaming, relational network topology diagrams, and event scene reconstruction diagrams. The teaching feedback module collects user interaction data on the visualization results and optimizes the weights of knowledge graph nodes through semantic similarity calculation, thereby realizing dynamic iteration of semantic analysis of historical teaching.

[0018] This invention constructs a full-process technical framework of "text standardization, semantic structuring, dynamic graph generation, and feedback iteration," achieving multi-dimensional technological innovation tailored to the specific characteristics and complexity of historical teaching scenarios. This significantly improves the accuracy, systematicity, and interactivity of semantic analysis in historical teaching, with the following specific benefits: I. Standardized processing ensures the semantic integrity of historical texts; By employing character-level cleaning and historical grammar rule verification, this invention effectively eliminates non-textual interference and grammatical errors, standardizes time representations and proper noun spellings, and ensures the semantic accuracy of the original text. Compared to traditional general text processing methods, this invention customizes a rule base specifically for the unique characteristics of historical documents (such as ancient Chinese sentence structures and reign-year dating methods), avoiding information distortion caused by inconsistent formatting or expression in historical semantics, and laying a high-quality data foundation for subsequent semantic analysis.

[0019] II. Multi-dimensional semantic extraction deepens the structuring of historical knowledge; By combining bidirectional LSTM and BERT models, this method achieves accurate extraction of historical event entities, interpersonal relationships, and timeline grammar, generating three-dimensional structured semantic units of "event-person-time". Breaking through the limitations of traditional keyword matching, it captures implicit semantic connections (such as causal relationships and temporal dependencies) in historical texts through deep learning models, transforming fragmented historical knowledge points into structured knowledge units and enhancing the interpretability and relevance of historical knowledge.

[0020] III. Dynamic knowledge graph construction enables visualization and association of historical knowledge; A dynamic knowledge graph is constructed based on structured semantic units, defining three-dimensional node attributes of events, figures, and time, as well as causal, temporal, and spatial edge relationships, to intuitively present the network structure of historical knowledge. The graph topology is optimized through graph neural networks to eliminate redundant information and strengthen key historical connections (such as the causal chain of "Opium War - Self-Strengthening Movement - Sino-Japanese War"), helping users quickly grasp the historical development and solving the problems of scattered and weakly connected historical knowledge in traditional teaching.

[0021] IV. Multimodal visualization enhances the interactive experience of history teaching; The knowledge graph is transformed into a timeline-based streaming display, a relational network topology diagram, and an event scene reconstruction diagram, presenting the results of historical semantic analysis in a multimodal format. The timeline supports chronological navigation of historical events, the topology diagram intuitively displays the relationships between people and events, and the scene reconstruction diagram enhances the perception of historical context, meeting the needs of different teaching scenarios (such as classroom teaching and self-study) and improving users' understanding and memory efficiency of historical knowledge.

[0022] V. Iterative optimization of semantic analysis accuracy based on teaching feedback; By collecting user interaction data and dynamically adjusting the weights of knowledge graph nodes using semantic similarity algorithms, the semantic analysis model can be continuously iterated. Based on user feedback, the accuracy of historical event recognition and relationship extraction is optimized, enabling the system to adapt to the learning preferences and cognitive needs of different users. This gradually improves the personalization and intelligence of semantic analysis in historical teaching, forming a closed-loop mechanism of "analysis-display-feedback-optimization".

[0023] In summary, this invention effectively solves the problems of complex text processing, inaccurate semantic extraction, and weak knowledge connections in history teaching through technological innovation. It provides history teaching with a semantic analysis tool that is accurate, systematic, and interactive, helping to improve the quality and efficiency of history teaching. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention have been clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0025] This embodiment is applied to a middle school history teaching system, and performs semantic analysis on the text of the textbook "Modern Chinese History" and extended historical materials (such as the original text of the Treaty of Nanjing, "Chronicle of Sun Yat-sen", etc.) to assist teachers in lesson preparation and students in self-study.

[0026] Step S1: Generate standardized historical text data; S11: Character-level cleaning. Input raw text data (such as OCR results of scanned textbooks, PDF historical texts), and remove non-text symbols using Python regular expressions: remove HTML tags (such as...). , Special symbols (such as #, @) and repeated character sequences (such as more than 3 consecutive ===); retain Chinese characters, English characters, numbers and historical proper noun symbols (such as the year "Guangxu 26th year", the sexagenary cycle "Gengzi"). Example: Clean the text "The Opium War broke out in 1840..." to "The Opium War broke out in 1840".

[0027] S12: Historical grammar rule verification, call the historical grammar rule library (built-in 3 types of core rules): I. Special sentence pattern rules in ancient Chinese. Judgment sentence structure: Identify the "……者……也” sentence pattern (such as "陈胜者,阳城人也”), generate the subject-predicate-object structure marker; match the "……者……也” sentence pattern through regular expressions (such as ([\u4e00-\u9fa5]+)者,([\u4e00-\u9fa5]+)也), extract the subject (such as "陈胜”) and the judgment object (such as "阳城人”), and generate the logical relationship of the subject-predicate-object structure. Passive sentence structure: Identify the "为……所……” sentence pattern (such as "若属皆且为所虏”), distinguish the agent and the patient, match the "为……所……” sentence pattern (such as 为([\u4e00-\u9fa5]+)所([\u4e00-\u9fa5]+)), distinguish the agent (such as "虏”) and the patient (such as "若属”), and mark the passive relationship. Ellipsis sentence supplement: For the common subject ellipsis in historical documents (such as "(诸葛亮)遂许先帝以驱驰”), infer and supplement the ellipsis through the context semantics. II. Historical document time expression rules. Year number chronology conversion: Convert "year number + year” to the Gregorian calendar (such as "光绪二十六年” → "1900年”, "民国元年” → "1912年”); Year number chronology conversion: Based on the year number-Gregorian calendar mapping table (such as {"光绪二十六年”:1900, "民国元年”:1912}), convert the year number in the text to the Gregorian calendar year through key-value pair matching; Map the Gregorian calendar year based on the 60-year cycle Heavenly Stems and Earthly Branches table (such as "甲午年” → "1894年”, "庚子年” → "1900年”). Fuzzy time definition: For interval time expressions such as "晚清” and "民国初期”, associate the specific Gregorian calendar range (such as "晚清” → "1840-1912年”). III. Historical proper noun spelling rules. Person name standardization: Verify the standard spelling of historical person names (such as "洪秀全” instead of "洪秀泉”, "曾国藩” instead of "曾国蕃”), and support the association of different names (such as "明太祖” → "朱元璋”). Geographical name evolution adaptation: Correct geographical names according to historical periods (such as "北平” → "北京” (after 1949), "金陵” → "南京”). Official position name standardization: Unify ancient official position terms (such as "丞相”, "巡抚”, "军机大臣”), and distinguish different positions with the same name (such as the different powers of "尚书” in different dynasties). Application method of the rule library. Grammar verification: Match each sentence of the cleaned historical text, and mark the grammar errors that do not conform to the above rules (such as "道光帝即位于1820年” → correct time expression; "洪秀泉起义” → wrong person name spelling). Semantic enhancement: Supplement the implicit semantics of historical texts through the rule library (such as "甲午战败” automatically associates with "1895年《马关条约》”), improve the integrity of subsequent semantic extraction; adopt the Heavenly Stems and Earthly Branches-Gregorian calendar comparison table (such as 甲午年=1894+60n), and lock the specific year in combination with the context of historical events (such as "甲午战败” → 1894年).The rule base is integrated into the system through structured storage (such as XML format) and supports dynamic updates (such as adding special sentence structures from unearthed documents). Output: Text with grammatical errors marked (e.g., "1911 Wuchang Uprising → grammatically correct" "Emperor Daoguang ascended the throne in 1820 → time statement correct"; "Chen Sheng was a man of Yangcheng → missing comma (error)" → corrected to "Chen Sheng was a man of Yangcheng").

[0028] S13: Standardized Format Conversion: Decompose the validated text into an "event-person-time" structure, extract core elements using the Semantic Role Labeling (SRL) algorithm, and unify the text format to an "event-person-time" triple structure: Time description: "Guangxu 26th year" → "1900", "Republic of China 10th year" → "1921"; Proper nouns: Terms such as "Three Principles of the People" and "Westernization Movement" are uniformly spelled according to the standard of the "Chinese Historical Dictionary"; Example: Standardize "January 1, 1912, Sun Yat-sen took office as Provisional President in Nanjing" to "Event: Founding of the Republic of China; Person: Sun Yat-sen; Time: 1912-01-01".

[0029] Step S2: Generation of structured historical semantic units; S21: Event Entity Recognition (Bidirectional LSTM Model). Input standardized text, and extract event entities using a pre-trained bidirectional LSTM model (training data includes 500,000 historical event samples). Input sentence: "In 1840, Britain launched the Opium War and attacked Guangzhou"; Output event triples: (Event name: Opium War, Time of occurrence: 1840, Location: Guangzhou); Entity type labels: Event (EVENT), Time (TIME), Location (LOCATION), with an accuracy of 92.3%.

[0030] S22: Character Relationship Extraction (BERT Pre-trained Model). This uses the BERT-base model (fine-tuned data includes 200,000 historical character relationship samples) to extract "character-action-object" triples. Input sentence: "Sun Yat-sen led the Xinhai Revolution, overthrowing the Qing Dynasty." Output relationship triples: (Character: Sun Yat-sen, Action: Lead, Object: Xinhai Revolution), (Character: Sun Yat-sen, Action: Overthrow, Object: Qing Dynasty). Relationship types: LEAD, COOPERATE, OPPOSE, etc., supporting 12 types of historical character relationships.

[0031] S23: Timeline Syntax Parsing (Dependency Parsing), using StanfordCoreNLP to parse time adverbial clauses and temporal relationships: Input sentence: "The May Fourth Movement occurred in 1919, and subsequently the Chinese Communist Party was founded"; Output timeline structure: May Fourth Movement (1919) → The influence of the magazine "New Youth" expanded (after 1919), with the temporal relationship marked as "after" (AFTER); Supports parsing complex time expressions, such as "ten years after the Opium War (1850)" and "during the First Sino-Japanese War (1894-1895)".

[0032] Step S3: Dynamic knowledge graph construction; S31: Node attribute definition (Neo4j graph database), Event nodes: include name (e.g., "Opium War"), type (war / reform / revolution), importance score (based on frequency of appearance in textbook, full score 10 points), and source (e.g., "People's Education Press, Required Course 1, p. 28"); Person nodes: include name, birth and death dates (e.g., "Sun Yat-sen, 1866-1925"), identity (politician / thinker), and faction (e.g., "revolutionary faction"); Time nodes: include timestamp (YYYY-MM-DD), dynasty / period (e.g., "late Qing Dynasty" "Republic of China").

[0033] S32: Edge relation weight calculation, causal association: weight is calculated based on the co-occurrence frequency of events, such as "Burning of Opium at Humen" and "Opium War" co-occurring 58 times, the weight is set to 0.8; temporal dependence: time interval decay function is used: weight = 1 / (time interval + 1), such as "Xinhai Revolution (1911)" and "May Fourth Movement (1919)" are 8 years apart, weight = 1 / (8+1)≈0.11; spatial mapping: calculated based on geographical distance, such as "Beijing" and "Tianjin" spatial weight is set to 0.9 (close distance), "Beijing" and "Guangzhou" are set to 0.3 (far distance).

[0034] S33: Graph structure optimization (Graph Neural Network GNN), using the GraphSAGE algorithm to eliminate isolated nodes (such as minor figures that appear only once) and redundant edges (such as weak relationships with correlation < 0.1); Example: Remove nodes without documentary evidence such as "Zhang San (participated in the Boxer Rebellion)," and retain the related edge "Empress Dowager Cixi - Boxer Rebellion" (weight 0.6).

[0035] Step S4: Multimodal visualization; S41: Streaming timeline display (implemented in D3.js), arranging event nodes in timestamp order. The horizontal axis represents the Gregorian calendar years (1840-1949), and the vertical axis represents the event type (war / reform / revolution). Interactive features: Supports mouse wheel zoom (enlarging the display of "details of 1911-1912") and range selection (selecting "1927-1937" to display the ten-year confrontation between the Kuomintang and the Communist Party). Example: The "1919" node on the timeline is labeled "May Fourth Movement," and clicking it expands the sub-events ("Burning of Zhao's House," "Removal of Cao Rulin"). Timestamp mapping and sorting: Based on the time attributes in the structured semantic units (such as "1840" and "1919"), event nodes are converted to Unix timestamps (such as "1840-01-01" → 1840*31536000 seconds), and event nodes are arranged in ascending order of timestamp. Supports historical periodization logic (e.g., "Late Qing Dynasty" and "Republic of China"), aggregating and displaying events from specific periods using a time interval algorithm (start_year ≤ event time ≤ end_year). Interactive zooming and interval selection calculation: The zoom function dynamically adjusts the timeline scale triggered by mouse wheel events (e.g., scale = scale * (1 ± 0.1)). Zooming in displays sub-events (e.g., expanding "May Fourth Movement" to "Burning of Zhao Family Mansion" and "Removal of Cao Rulin"), while zooming out merges similar events (e.g., merging "Opium War-related Treaties"). Interval selection calculates the time range by selecting coordinates with the mouse (x1 ≤ event timestamp ≤ x2), filtering and highlighting all event nodes within that interval.

[0036] S42: Relationship network topology graph (Force-Directed algorithm). Node size maps to event importance (e.g., "Opium War" node diameter 20px, "Second Opium War" 15px); edge thickness maps to relationship weight (e.g., "Self-Strengthening Movement - Sino-Japanese War" causal edge width 5px, "Hundred Days' Reform - Boxer Rebellion" chronological edge width 2px); layout: character nodes (blue), event nodes (red), time nodes (gray), automatically avoiding node overlap through the force-directed algorithm. Force-directed layout algorithm implementation: Node repulsion calculation: using the Fruchterman-Reingold algorithm, the repulsion formula between nodes is F_rep = k² / d (k is the layout coefficient, d is the node distance), avoiding character / event node overlap. Edge attraction calculation: the attraction formula for character-event related edges is F_attr = d² / k, edges with higher weights (e.g., "Sun Yat-sen - Xinhai Revolution", weight 0.9) have stronger attraction, and the nodes are closer together. Visual mapping of nodes and edges. Node size calculation: Event node diameter size = base_size * (1 + importance score / 10) (base_size = 10px, importance score is based on frequency of occurrence in the textbook, with a maximum score of 10). For example, the diameter of "Opium War" (score 9) is 10 * (1 + 0.9) = 19px. Edge thickness calculation: Width of related edges width = base_width * weight (base_width = 2px). For example, the width of a causal related edge (weight 0.8) is 2 * 0.8 = 1.6px, and the width of a temporal dependent edge (weight 0.3) is 0.6px.

[0037] S43: Event Scene Reconstruction Image (GAN Model Generation). Input event entities and attributes (e.g., "Burning of Opium at Humen, 1839, Lin Zexu, Guangzhou"), and call the StyleGAN2 model to generate a semi-realistic image: Character Image: Generated based on historical portraits (e.g., portrait of Lin Zexu in official robes) and a Qing Dynasty clothing database; Scene Elements: Integrates geographical data of the Humen Fort in Guangzhou, 19th-century sailboat models, etc.; Labeling: Labels the event name, time, and key figures below the image (e.g., "Burning of Opium at Humen (1839) - Lin Zexu presided over the anti-opium movement"). Historical scene material matching and entity attribute extraction: Extract era characteristics (e.g., "1839" → Qing Dynasty), regional style (e.g., "Guangzhou" → Lingnan architecture), and character attributes (e.g., "Lin Zexu" → Qing Dynasty official clothing); Material library retrieval: Based on attribute keywords ("Qing Dynasty + Guangzhou + official + fort"), call the historical scene material library and match resources such as the Qing Dynasty clothing database and the 3D model of the Humen Fort in Guangzhou. Image synthesis using GAN networks: The StyleGAN2 model input encodes event entity attributes (time, people, location) into conditional vectors, which are then input into the pre-trained StyleGAN2 model to generate semi-realistic images. Historical research constraints are applied: image details are corrected using historical clothing databases (such as Qing Dynasty official robe patterns and official hat / feather specifications) to ensure that clothing and objects conform to historical facts (e.g., the "peacock patch" pattern on Lin Zexu's official robe). Event element annotation calculation: Based on the image semantic segmentation results, text boxes are generated below the core area of ​​the person / event (e.g., "Burning of Opium at Humen (1839) - Lin Zexu presided over the anti-opium movement"), with font size proportional to image size (font_size = image_height / 50). Through these calculation methods, the timeline, topology map, and scene reconstruction map achieve multimodal visualization of historical knowledge, enhancing the contextual awareness and interactive experience of historical teaching.

[0038] Step S5: Teaching feedback and dynamic iteration; S51: Interactive data collection, recording user behavior: click event nodes (e.g., "Xinhai Revolution" clicks 200 times), dwell time (e.g., "May Fourth Movement" average dwell time 45 seconds), path trajectory (e.g., "Opium War → Self-Strengthening Movement → Sino-Japanese War" browsing sequence); data storage: using MongoDB database to store user ID, interaction time, operation type (click / zoom / download).

[0039] S52: Semantic similarity optimization, calculating the cosine similarity between user interaction data and knowledge graph nodes: Example: If the user frequently clicks on nodes related to "Sun Yat-sen", calculate its similarity with "Xinhai Revolution" and "Three Principles of the People" (0.92, 0.88), and increase the weights of the corresponding nodes; Adjust the display priority:置顶 display the nodes with high similarity (such as "Sun Yat-sen - Xinhai Revolution") on the timeline. Optimize the weights of knowledge graph nodes and iteratively update the semantic analysis model. By calculating the semantic similarity between user interaction feedback data and knowledge graph nodes, adjust the node weights and display priorities. Input data: User interaction feedback matrix (including behavioral data such as click heat, stay duration, and path trajectory). Output result: Optimized node weights (such as the weight of the "Sun Yat-sen" node frequently visited by users is increased).

[0040] S53: Model iterative update, optimizing the semantic extraction model through the reinforcement learning algorithm (PPO): Reward function: Based on the user click accuracy (such as if the model recognizes "Reform Movement of 1898" and is clicked by the user, the reward is +1); Iteration period: Update the bidirectional LSTM event recognition model every two weeks, increasing the accuracy from 92.3% to 94.1%. Iterative update of the semantic analysis model. Based on the semantic similarity results, update the semantic extraction model (such as the bidirectional LSTM event recognition model) through the reinforcement learning algorithm to improve the historical event recognition accuracy. Use the cosine similarity algorithm to achieve semantic matching. Specific steps: Feature vector construction: Convert knowledge graph nodes (events, people, times) into high-dimensional semantic vectors (such as training historical proper nouns embedding through the Word2Vec model). Convert user interaction data (such as click sequences, stay durations) into behavioral feature vectors, with weight assignments as follows: click heat (0.4), stay duration (0.3), path continuity (0.3). Calculate the cosine similarity through the formula: ; where A (knowledge graph node vector) and B (user behavior feature vector) are the core vector parameters, and n (vector dimension) is the boundary condition for the summation calculation. Specific steps: Feature vector construction: Convert knowledge graph nodes (events, people, times) into high-dimensional semantic vectors A (such as training historical proper nouns embedding through the Word2Vec model); Convert user interaction data (such as click sequences, stay durations) into behavioral feature vectors B, with weight assignments as follows: click heat (0.4), stay duration (0.3), path continuity (0.3).

[0041] Cosine similarity calculation: Calculate the similarity between vector A and B through the following formula. In the formula, n represents the dimensions of vector A and B (the dimensions of the two need to be the same), where: A: High-dimensional semantic vector of knowledge graph nodes (such as the semantic vector of the "Sun Yat-sen" node); B: Feature vector of user interaction behavior (such as the behavior vector of users clicking on "Sun Yat-sen" or "Xinhai Revolution"); n: The dimensions of vectors A and B (e.g., a vector trained based on a 500-dimensional historical term vocabulary, n=500). Ai and Bi: These are the components of vectors A and B in the i-th dimension, respectively.

[0042] Example: Calculate the similarity between the "Sun Yat-sen" node vector and the user click sequence vector. If the result is 0.92 (close to 1), it is determined that the user has a high level of attention to this node, thus increasing its display priority in the knowledge graph. Threshold determination and weight adjustment: Set a similarity threshold (e.g., 0.7). Nodes with weights above the threshold are weighted (e.g., +0.2), while nodes with weights below the threshold are weighted (e.g., -0.1), achieving dynamic optimization. Integration of historical semantic features: Unlike the general cosine similarity algorithm, this invention incorporates historical proper noun embeddings (e.g., domain word vectors for "Westernization Movement" and "Three Principles of the People") during the vector construction stage, improving the accuracy of similarity calculation in historical scenarios. Multi-dimensional behavioral weighting: The weight allocation of user interaction data (click popularity > dwell time > path continuity) is customized for historical teaching scenarios, strengthening the identification of attention to core events and key figures.

[0043] Through the above methods, the semantic similarity algorithm achieves dynamic adaptation between user interaction feedback and knowledge graph structure, thereby improving the personalization and accuracy of semantic analysis in history teaching.

[0044] This embodiment achieves deep semantic analysis of historical teaching texts through standardized text processing, multi-dimensional semantic extraction, dynamic knowledge graph construction, and visual feedback. In this embodiment, processing the text of the "Modern Chinese History" textbook (50,000 words) took 8 minutes, generating a knowledge graph containing 237 event nodes, 189 character nodes, and 526 relationship edges. Students' accuracy in associating historical events with memory improved by 38%, and teachers' lesson preparation efficiency improved by 45%.

[0045] Note: The models (such as bidirectional LSTM, BERT), databases (Neo4j, MongoDB), and visualization tools (D3.js) involved in this embodiment are all existing technologies. All steps can be implemented using a general-purpose computer (Python, Neo4j, MongoDB, D3.js) without the need for special equipment. The specific implementation details do not affect the inventiveness of this invention. The core innovation lies in the design of a full-process method for "text standardization - semantic structuring - graph dynamization - feedback iteration" in the context of historical teaching.

[0046] To more clearly explain this invention, the semantic analysis method based on history teaching has the following flowchart framework: It adopts a linear progressive + feedback closed-loop structure, which includes 5 main steps (S1-S5). Each step is further refined through sub-steps, and finally forms a dynamic iterative closed loop through the teaching feedback module.

[0047] Main steps and relationships: Standardized historical text data generation (S1); Input: Original history teaching text data (textbooks, historical materials, etc.).

[0048] Sub-steps: S11: Character-level cleaning (removal of non-text symbols and repeating sequences); S12: Historical grammar rule verification (verification of classical Chinese sentence structure, time expression, and proper nouns). S13: Standardized format conversion (unified time representation, proper noun spelling).

[0049] Output: Standardized historical text data.

[0050] Flow: S1→S2 (Standardized text as input for semantic extraction).

[0051] Structured historical semantic unit generation (S2); Input: Standardized historical text data.

[0052] Sub-steps: S21: Event Entity Recognition (Bidirectional LSTM model extracts event name, time, and location); S22: Character Relationship Extraction (BERT model generates "character-action-object" triples); S23: Timeline syntax parsing (dependency syntax analysis of causal and temporal relationships).

[0053] Output: Structured historical semantic units (three-dimensional data of events, people, and time).

[0054] Flow direction: S2→S3 (semantic units serve as the foundation for knowledge graph construction).

[0055] Dynamic knowledge graph construction (S3); Input: Structured historical semantic units.

[0056] Sub-steps: S31: Node attribute definition (attributes such as type and importance of event, person, and time node); S32: Edge relation weight calculation (weight assignment for causal, temporal, and spatial relations); S33: Graph structure optimization (graph neural network to eliminate isolated nodes and redundant edges).

[0057] Output: Dynamic historical knowledge graph (including the network structure of node-edge relationships).

[0058] Flow: S3→S4 (Knowledge graph as the data source for visualization).

[0059] Multimodal visualization (S4); Input: Dynamic historical knowledge graph.

[0060] Sub-steps: S41: Timeline flow display (events are arranged by timestamp, and zoom and range selection are supported); S42: Relationship network topology diagram (force-oriented layout showing the relationship between people and events); S43: Event scene reconstruction image (semi-realistic historical scene generated by GAN model).

[0061] Output: Multimodal visualization results (timeline, topology diagram, scene diagram).

[0062] Flow: S4→S5 (Visualized results for user interaction, generating feedback data).

[0063] Teaching feedback and iterative optimization (S5); Input: User interaction feedback data (number of clicks, time spent on the page, browsing path).

[0064] Sub-steps: S51: Interactive data collection (recording and storing user behavior); S52: Semantic similarity optimization (calculate the similarity between user preferences and nodes, and adjust the weights); S53: Model Iterative Update (Reinforcement Learning to Optimize the Semantic Extraction Model).

[0065] Output: Optimized semantic analysis model and knowledge graph.

[0066] Flow: S5→S2 (The optimized model is used for the next round of semantic extraction, forming a closed loop).

[0067] Core logical relationship: Progressive relationship: S1 (data standardization) → S2 (semantic extraction) → S3 (graph construction) → S4 (visualization), realizing the whole process from raw text to structured knowledge and then to multimodal display; Closed-loop relationship: S4 (user interaction) → S5 (feedback optimization) → S2 (model update), iteratively optimizes semantic extraction accuracy through user behavior data, and dynamically adapts to teaching needs; Dependencies: The sub-steps of each step are expanded according to the logic of "preprocessing → core processing → optimization" (such as S11 → S12 → S13) to ensure the accuracy and availability of the output data at each step.

[0068] Through the above structure and process, the entire life cycle of semantic analysis in historical teaching, namely "data input-processing-output-feedback-iteration", is fully presented, highlighting the core technical path of standardized processing of historical texts, multi-dimensional semantic extraction, dynamic knowledge organization and teaching adaptability optimization.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A semantic analysis method based on history teaching, characterized in that, It includes the following steps: S1: Obtain historical teaching text data, perform character-level preprocessing and grammar rule verification on the text data, and generate standardized historical text data; S2: Perform semantic extraction on the standardized historical text data, identify historical event entities, character relationships, and timeline grammar structures, and generate structured historical semantic units; S3: Build a dynamic knowledge graph based on the structured historical semantic units, and define the three-dimensional node attributes of event-character-time and the edge relationships of causality-temporal sequence-space; S4: Convert the dynamic knowledge graph into a multi-modal visualization result, and present the semantic analysis result through a time-axis streaming display, a relational network topology diagram, and an event scene restoration diagram; S5: Collect user interaction feedback data, optimize the node weights of the knowledge graph through a semantic similarity algorithm, and iteratively update the semantic analysis model.

2. The method according to claim 1, characterized in that, Step S1 includes the following steps: S11: Perform character-level cleaning on the historical teaching text data, and remove non-text symbols and duplicate character sequences; S12: Perform grammar verification on the cleaned text based on the historical grammar rule library, and mark the subject-predicate-object structure and adverbial clauses of time; S13: Convert the text after grammar verification into a standardized format, and unify the time expression and the spelling of historical proper nouns.

3. The method according to claim 1, characterized in that, Step S2 includes the following steps: S21: Use a bidirectional LSTM model to perform event entity recognition on the standardized historical text data, and extract event names, occurrence times, and location information; S22: Extract character entities and relationship triples through a BERT pre-trained model, and generate "character-action-object" semantic units; S23: Based on dependency syntax analysis of the timeline grammar structure, establish the mapping relationship of the event occurrence order and time interval.

4. The method according to claim 1, characterized in that, Step S3 includes the following steps: S31: Define the node attributes of the dynamic knowledge graph. The event node includes a type label, an importance score, and a literature source; S32: Build an edge relationship weight calculation model. The causal association weight is based on the event co-occurrence frequency, and the temporal dependence weight is based on the time interval decay function; S33: Optimize the structure of the knowledge graph through a graph neural network, and eliminate isolated nodes and redundant edge relationships.

5. The method according to claim 1, characterized in that, Step S4 includes the following steps: S41: The time-axis streaming display module arranges event nodes in chronological order, and supports zooming and interval selection interactions; S42: The relational network topology diagram uses a force-directed layout algorithm. The node size maps the event importance, and the edge thickness maps the relationship weight; S43: The event scene restoration diagram generates a semi-realistic image in combination with entity attributes, and marks key characters and event elements.

6. The method according to claim 1, characterized in that, Step S5 includes the following steps: S51: Collect user click heat, stay duration, and path trajectory data, and generate an interaction feedback matrix; S52: Calculate the semantic similarity between the feedback matrix and the nodes of the knowledge graph, and adjust the node weights and display priorities; S53: Iteratively update the semantic extraction model through a reinforcement learning algorithm, and optimize the historical event recognition accuracy.

7. The method according to claim 2, characterized in that, The historical grammar rule library in step S12 includes: S121: Rules for special sentence patterns in ancient Chinese, including the "……者……也" structure of judgment sentences and the "为……所……" structure of passive sentences; S122: Rules for expressing time in historical documents, covering the automatic conversion between reign titles, sexagenary cycle, and the Gregorian calendar; S123: Spelling rules for historical proper nouns, including alternative names of people, historical changes of place names, and standardization of official titles.

8. The method according to claim 3, characterized in that, The relation triple extraction in step S22 includes: S221: Person-Event Association, marking the role and degree of influence of a person in an event; S222: Character-Character Relationships, distinguishing between kinship, superior-subordinate relationships, and cooperative / antagonistic relationships; S223: Event-Location Association, records the precise geographical location and spatial extent of the event.

9. The method according to claim 5, characterized in that, The generation of the event scene reconstruction map in step S43 includes: S431: Call the historical scene material library based on event node attributes to match era characteristics and regional styles; S432: Semi-realistic images are synthesized using a GAN network, with figures' clothing and objects conforming to historical research. S433: Overlay event element annotations, including timestamps, names of key figures, and summaries of core events.

10. A semantic analysis system based on history teaching, used to perform the method as described in any one of claims 1 to 9, characterized in that, include: The data processing module is used to acquire historical teaching text data, perform character-level cleaning and grammar rule matching on the text data, and generate standardized historical text data. The semantic analysis module performs multi-dimensional semantic extraction on standardized historical text data, including historical event entity recognition, person relationship extraction, and timeline grammar parsing, to generate structured historical semantic units. The knowledge graph construction module builds a dynamic knowledge graph based on structured historical semantic units. Graph nodes contain event types, person attributes, and timestamp information, and edge relationships include causal associations, temporal dependencies, and spatial mappings. The image display module converts dynamic knowledge graphs into multimodal visualization results, including timeline streaming, relational network topology diagrams, and event scene reconstruction diagrams. The teaching feedback module collects user interaction data on the visualization results and optimizes the weights of knowledge graph nodes through semantic similarity calculation, thereby realizing dynamic iteration of semantic analysis of historical teaching.